1. Index Compression David Kauchak cs458 Fall 2012 adapted from: http://www.stanford.edu/class/cs276/handouts/lecture5-indexcompression.ppt

2. Administrative Assignment 1? Homework 2 out “What I did last summer” lunch talks today

4. Distributed indexing Maintain a master machine directing the indexing job Break up indexing into sets of (parallel) tasks Master machine assigns each task to an idle machine from a pool Besides speed, one advantage of a distributed scheme is fault tolerance Master Tasks

5. Distributed indexing Quick refresh of the non-parallelized approach: I did enact Julius Caesar I was killed i' the Capitol; Brutus killed me. So let it be with Caesar. The noble Brutus hath told you Caesar was ambitious word 1 word 2 word n … 1. create term list 2. sort term list 3. create postings list

6. Distributed indexing I did enact Julius Caesar I was killed i' the Capitol; Brutus killed me. So let it be with Caesar. The noble Brutus hath told you Caesar was ambitious word 1 word 2 word n … 1. create term list 2. sort term list 3. create postings list Split into smaller, parallelizable chunks

7. Parallel tasks We will use two sets of parallel tasks Parsers (Step 1: create term list) Inverters (Steps 2-3: sort term list, create postings list) split documents up for parsers all docs

8. Parsers Read a document at a time and emits (term, doc) pairs (Step 1) Parser writes pairs into j partitions Each partition is for a range of terms’ first letters (e.g., a-f, g-p, q-z) – here j=3. a-f g-p q-z did1 enact1 caesar1 capitol1 brutus1 be2 caesar2 brutus2 caesar2 was1 the1 so2 with2 told2 was2

9. Inverters Collects all (term, doc) pairs for one term-partition Sorts and writes to postings lists a-f index for a-f a-f 2. sort term list3. create postings list1b. concatenate

10. Data flow splits Parser Master a-fg-pq-z a-fg-pq-z a-fg-pq-z Inverter Postings a-f g-p q-z assign Map phase Segment files Reduce phase

11. MapReduce MapReduce (Dean and Ghemawat 2004) is a robust and simple framework for distributed computing without having to write code for the distribution part The Google indexing system (ca. 2002) consists of a number of phases, each implemented in MapReduce MapReduce and similar type setups are hugely popular for web-scale development!

12. MapReduce Index construction is just one phase After indexing, we need to be ready to answer queries There are two ways to we can partition the index: Term-partitioned: one machine handles a subrange of terms Document-partitioned: one machine handles a subrange of documents Which do you think search engines use? Why? word 1 word 2 word n …

13. Index compression Compression techniques attempt to decrease the space required to store an index What other benefits does compression have? Keep more stuff in memory (increases speed) Increase data transfer from disk to memory [read compressed data and decompress] is faster than [read uncompressed data] What does this assume? Decompression algorithms are fast True of the decompression algorithms we use

14. How does the vocabulary size grow with the size of the corpus? number of documents vocabulary size

15. How does the vocabulary size grow with the size of the corpus? log of the number of documents log of the vocabulary size

16. Heaps’ law Does this explain the plot we saw before? What does this say about the vocabulary size as we increase the number of documents? there are almost always new words to be seen: increasing the number of documents increases the vocabulary size to get a linear increase in vocab size, need to add exponential number of documents vocab size = k (tokens) b V = k T b log V= log k + b log(T) Typical values: 30 ≤ k ≤ 100 b ≈ 0.5

17. vocab growth vs. size of the corpus log of the number of documents log of the vocabulary size log 10 M = 0.49 log 10 T + 1.64 is the best least squares fit. M = 10 1.64 T 0.49 k = 10 1.64 ≈ 44 b = 0.49.

18. Discussion How do token normalization techniques and similar efforts like spelling correction interact with Heaps’ law? vocab size = k (tokens) b V = k T b Typical values: 30 ≤ k ≤ 100 b ≈ 0.5

19. Heaps’ law and compression index compression is the task of reducing the memory requirement for storing the index What implications does Heaps’ law have for compression? Dictionary sizes will continue to increase Dictionaries can be very large

20. How does a word’s frequency relate to it’s frequency rank word’s frequency rank word frequency

21. How does a word’s frequency relate to it’s frequency rank log of the frequency rank log of the frequency

22. Zipf’s law In natural language, there are a few very frequent terms and very many very rare terms Zipf’s law: The i th most frequent term has frequency proportional to 1/i where c is a constant frequency i ∝ c/i log(frequency i ) ∝ log c – log i

23. Consequences of Zipf’s law If the most frequent term (the) occurs cf 1 times, how often do the 2 nd and 3 rd most frequent occur? then the second most frequent term (of) occurs cf 1 /2 times the third most frequent term (and) occurs cf 1 /3 times … If we’re counting the number of words in a given frequency range, lowering the frequency band linearly results in an exponential increase in the number of words

24. Zipf’s law and compression What implications does Zipf’s law have for compression? word’s frequency rank word frequency Some terms will occur very frequently in positional postings lists Dealing with these well can drastically reduce the index size

25. Compresssing the inverted index word 1 word 2 word n … What do we need to store? How are we storing it?

26. Compressing the inverted index Two things to worry about: dictionary: make it small enough to keep in main memory Search begins with the dictionary postings: Reduce disk space needed, decrease time to read from disk Large search engines keep a significant part of postings in memory

27. Lossless vs. lossy compression What is the difference between lossy and lossless compression techniques? Lossless compression: All information is preserved Lossy compression: Discard some information, but attempt to keep information that is relevant Several of the preprocessing steps can be viewed as lossy compression: case folding, stop words, stemming, number elimination. Prune postings entries that are unlikely to turn up in the top k list for any query Where else have you seen lossy and lossless compresion techniques?

28. The dictionary If I asked you to implement it right now, how would you do it? How much memory would this use? word 1 word 2 word n …

29. The dictionary Array of fixed-width entries ~400K terms; 28 bytes/term = 11.2 MB. 20 bytes4 bytes each (assuming 32-bit)(assume 1byte chars)

30. Fixed-width terms are wasteful Any problems with this approach? Most of the bytes in the Term column are wasted – we allocate 20 bytes for 1 letter terms And we still can’t handle supercalifragilisticexpialidocious Written English averages ~4.5 characters/word Is this the number to use for estimating the dictionary size? Ave. dictionary word in English: ~8 characters Short words dominate token counts but not type average

31. Any ideas? Store the dictionary as one long string Gets ride of wasted space If the average word is 8 characters, what is our savings over the 20 byte representation? Theoretically, 60% Any issues? ….systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo….

32. Dictionary-as-a-String Store dictionary as a (long) string of characters: Pointer to next word shows end of current word ….systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo…. Total string length = 400K x 8B = 3.2MB Pointers resolve 3.2M positions: log 2 3.2M = 22bits = 3bytes How much memory to store the pointers?

33. Space for dictionary as a string Fixed-width 20 bytes per term = 8 MB As a string 5.6 MB (3.2 for dictionary and 2.4 for pointers) 30% reduction! Still a long way from 60%. Any way we can store less pointers?

34. Blocking Store pointers to every kth term string ….systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo…. What else do we need?

35. Blocking Store pointers to every kth term string Example below: k = 4 Need to store term lengths (1 extra byte) ….7systile9syzygetic8syzygial6syzygy11szaibelyite8szczecin9szomo…. Save 9 bytes  on 3  pointers. Lose 4 bytes on term lengths.

36. Net Where we used 3 bytes/pointer without blocking 3 x 4 = 12 bytes for k=4 pointers, now we use 3+4=7 bytes for 4 pointers. Shaved another ~0.5MB; can save more with larger k. Why not go with larger k?

37. Dictionary search without blocking How would we search for a dictionary entry? ….systilesyzygeticsyzygialsyzygyszaibelyiteszczecinszomo….

38. Dictionary search without blocking Binary search Assuming each dictionary term is equally likely in query (not really so in practice!), average number of comparisons = ? (1 + 2*2+4*3+4)/8 = 2.6

39. Dictionary search with blocking What about with blocking? ….7systile9syzygetic8syzygial6syzygy11szaibelyite8szczecin9szomo….

40. Dictionary search with blocking Binary search down to 4-term block Then linear search through terms in block. Blocks of 4 (binary tree), avg. = ? (1+2∙2+2∙3+2∙4+5)/8 = 3 compares

41. More improvements… We’re storing the words in sorted order Any way that we could further compress this block? 8automata8automate9automatic10automation

42. Front coding Front-coding: Sorted words commonly have long common prefixes – store differences only (for last k-1 in a block of k) 8automata8automate9automatic10automation  8automat*a1e2ic3ion Encodes automat Extra length beyond automat Begins to resemble general string compression

43. RCV1 dictionary compression TechniqueSize in MB Fixed width11.2 String with pointers to every term7.6 Blocking k = 47.1 Blocking + front coding5.9

44. Postings compression The postings file is much larger than the dictionary, by a factor of at least 10 A posting for our purposes is a docID Regardless of our postings list data structure, we need to store all of the docIDs For Reuters (800,000 documents), we would use 32 bits per docID when using 4-byte integers Alternatively, we can use log 2 800,000 ≈ 20 bits per docID

45. Postings: two conflicting forces Where is most of the storage going? Frequent terms will occur in most of the documents and require a lot of space A term like the occurs in virtually every doc, so 20 bits/posting is too expensive. Prefer 0/1 bitmap vector in this case A term like arachnocentric occurs in maybe one doc out of a million – we would like to store this posting using log 2 1M ~ 20 bits.

46. Postings file entry We store the list of docs containing a term in increasing order of docID. computer: 33,47,154,159,202 … Is there another way we could store this sorted data? Store gaps: 33,14,107,5,43 … 14 = 47-33 107 = 154 – 47 5 = 159 - 154

47. Fixed-width How many bits do we need to encode the gaps? Does this buy us anything?

48. Variable length encoding Aim: For arachnocentric, we will use ~20 bits/gap entry For the, we will use ~1 bit/gap entry Key challenge: encode every integer (gap) with as few bits as needed for that integer 1, 5, 5000, 1, 1524723, … for smaller integers, use fewer bits for larger integers, use more bits

49. Variable length coding 1, 5, 5000, 1, 1124 … 1, 101, 1001110001, 1, 10001100101 … Fixed width: 000000000100000001011001110001 … every 10 bits Variable width: 11011001110001110001100101 … ?

50. Variable Byte (VB) codes Rather than use 20 bits, i.e. record gaps with the smallest number of bytes to store the gap 1, 101, 1001110001 00000001, 00000101, 00000010 01110001 1 byte 2 bytes 00000001000001010000001001110001 ?

51. VB codes Reserve the first bit of each byte as the continuation bit If the bit is 1, then we’re at the end of the bytes for the gap If the bit is 0, there are more bytes to read For each byte used, how many bits of the gap are we storing? 1, 101, 1001110001 100000011000010100000100 11110001

52. Example docIDs824829215406 gaps5214577 VB code00000110 10111000 1000010100001101 00001100 10110001 Postings stored as the byte concatenation 000001101011100010000101000011010000110010110001 Key property: VB-encoded postings are uniquely prefix-decodable. For a small gap (5), VB uses a whole byte.

53. Other variable codes Instead of bytes, we can also use a different “unit of alignment”: 32 bits (words), 16 bits, 4 bits (nibbles) etc. What are the pros/cons of a smaller/larger unit of alignment? Larger units waste less space on continuation bits (1 of 32 vs. 1 of 8) Smaller unites waste less space on encoding smaller number, e.g. to encode ‘1’ we waste (6 bits vs. 30 bits)

54. More codes Still seems wasteful What is the major challenge for these variable length codes? We need to know the length of the number! Idea: Encode the length of the number so that we know how many bits to read 100000011000010100000100 11110001

55. Gamma codes Represent a gap as a pair length and offset offset is G in binary, with the leading bit cut off 13 → 1101 → 101 17 → 10001 → 0001 50 → 110010 → 10010 length is the length of offset 13 (offset 101), it is 3 17 (offset 0001), it is 4 50 (offset 10010), it is 5

56. Encoding the length We’ve stated what the length is, but not how to encode it What is a requirement of our length encoding? Lengths will have variable length (e.g. 3, 4, 5 bits) We must be able to decode it without any ambiguity Any ideas? Unary code Encode a number n as n 1’s, followed by a 0, to mark the end of it 5 → 111110 12 → 1111111111110

57. Gamma code examples numberlengthoffset  -code 0 1 2 3 4 9 13 24 511 1025

58. Gamma code examples numberlengthoffset  -code 0none 100 210010,0 310110,1 411000110,00 911100011110,001 1311101011110,101 2411110100011110,1000 51111111111011111111111111110,11111111 102511111111110000000000111111111110,0000000001

59. Gamma code properties Uniquely prefix-decodable, like VB All gamma codes have an odd number of bits What is the fewest number of bits we could expect to express a gap (without any other knowledge of the other gaps)? log 2 (gap) How many bits do gamma codes use? 2  log 2 (gap)  +1 bits Almost within a factor of 2 of best possible

60. Gamma seldom used in practice Machines have word boundaries – 8, 16, 32 bits Compressing and manipulating at individual bit- granularity will slow down query processing Variable byte alignment is potentially more efficient Regardless of efficiency, variable byte is conceptually simpler at little additional space cost

61. RCV1 compression Data structureSize in MB dictionary, fixed-width11.2 dictionary, term pointers into string7.6 with blocking, k = 47.1 with blocking & front coding5.9 collection (text, xml markup etc)3,600.0 collection (text)960.0 Term-doc incidence matrix40,000.0 postings, uncompressed (32-bit words)400.0 postings, uncompressed (20 bits)250.0 postings, variable byte encoded116.0 postings,  encoded 101.0

62. Index compression summary We can now create an index for highly efficient Boolean retrieval that is very space efficient Only 4% of the total size of the collection Only 10-15% of the total size of the text in the collection However, we’ve ignored positional information Hence, space savings are less for indexes used in practice But techniques substantially the same